Evidence hubLedger current to 18 Oct 2026 · recounted 18 Oct 2026 · claims rechecked 18 Oct 2026, 05:19 UTC · jkudish/jev-mcp read and check commits shown separately, and Vercel's production-index figures added beside the relayed Vercel claim, 19 Oct 2026 · repositories rechecked 3 Oct 2026, 05:23 UTC

Jev examples: products and projects built with Jev (TypeSafe AI), graded by evidence

This is an evidence ledger, not a customer list. Each project is graded by what the pool could check: whether it read the code that calls Jev, only the project's description, or a mocked demo. Google, YouTube, Bing and DuckDuckGo all suggested "jev examples" in autocomplete on 3 Oct 2026, 05:22 UTC; none of them suggested "jev products". This page answers that search with graded rows rather than a showcase.

Key finding

Of 113 Jev products and projects in the ledger, the call to Jev was read in code at a pinned commit for 100; none is independently verified as running in production.

Documentedledger counts, recounted 18 Oct 2026Reported7 published claims, rechecked 18 Oct 2026, 05:19 UTC (5 of 7; 2 not rechecked)

Grade A means the pool read the code path that calls Jev at a stated commit. It does not mean the project runs, has users or works well.

Short answer

As of 18 October 2026 the ledger has 113 rows (recounted 18 Oct 2026; 99 on 11 Oct). For 100 of 113 the pool read the code that calls Jev at a pinned commit (grade A); 12 of 113 rest on the projects' own descriptions (B); 1 of 113 is a mock (C). Independently verified in production: 0 of 113. Seven published production or cost claims are listed with what their sources show. None is independently verified in production. This does not mean none exists.

The rows include a Gmail sorter, a WordPress moderation plugin, a trading gate, SQL extensions, browser agents, Claude Code, MCP and agent-guard integrations, and, added on 3 Oct, 24 ticket-triage, model-routing, moderation, email and n8n projects. On 4 Oct, 6 trading, fraud, expense-approval and judging projects were added and 2 trading bots moved from B to A. On 5 Oct, 5 browser and computer-use projects were added and 3 moved from B to A. On 6 Oct, 18 coding-agent integrations (Codex, OpenCode, Cursor, Hermes Agent, OpenClaw) and 2 Claude Code guard rows were added, all grade A; the selection is the rows read for build/coding-agents plus the two guard rows, not a survey of coding-agent plugins. On 8 Oct, 2 rows were added (LanceDB TypeSafeReranker and pg-redact) and jev-reranker (Python) moved from B to A after its v0.1.2 code was read; see reranking and RAG and Jev inside SQL and document pipelines. On 10 Oct, 2 more coding-agent integrations were added, both grade A: jev-effort-router (a Hermes Agent plugin that sets the model and reasoning effort per turn) and jev-gateway (an npm proxy in which Jev picks the tool for Codex, OpenCode, Claude Code, Kilo, Gemini and Devin). On 11 Oct, 3 more coding-agent integrations were added, all grade A: jev-triage (an OpenClaw plugin that can answer or drop an inbound message without the main model), the jev-for-all start launcher (Jev picks the model tier and effort for Claude Code or OpenCode) and hermes-jev-mana (a Hermes Agent core patch, inert after its documented install). On 18 Oct, 14 rows were added, all grade A: three Claude Code guards (claude-model-guard, claude-jev-advisor and ohmyjev), three game and robot demos (doom-jev, jev-robotics-demo and JevPlaysSnakeGame, compared on real-time demos) and eight projects; see Added 18 Oct 2026. Two of the 113 are official TypeSafe: the Claude Code plugin, which makes no Jev call, and the n8n node. One of the 113, @openclaw/typesafe, is published by a host vendor (OpenClaw), not by TypeSafe.

Ledger rows by evidence grade, as of 18 October 2026

A · call path read15 through 29 Sep, 24 added 3 Oct, 8 added or upgraded 4 Oct, 8 added or upgraded 5 Oct, 20 added 6 Oct, 6 added or upgraded 8 Oct, 2 added 10 Oct, 3 added 11 Oct, 14 added 18 Oct100 of 113
B · project docs describe Jev use2 moved to A on 4 Oct, 3 on 5 Oct, 4 on 8 Oct12 of 113
C · scripted or mocked1 of 113
Independently verified in production0 of 113

113 rows by primary grade (recounted 18 Oct 2026). In the A bar, from left to right: the 15 rows graded A through 29 Sep, the 24 added on 3 Oct, the 8 added or upgraded on 4 Oct (6 new rows, 2 moved from B), the 8 added or upgraded on 5 Oct (5 new rows, 3 moved from B; outlined), the 20 added on 6 Oct (dashed outline), the 6 added or upgraded on 8 Oct (2 new rows, 4 moved from B; third-to-last segment), the 2 added on 10 Oct, the 3 added on 11 Oct (second-to-last segment) and the 14 added on 18 Oct (last segment, solid outline). Bar lengths use one scale for all grades. Mixed grades (for example B/C) are counted under their first letter. No project in any grade has an independently verified production deployment. Documented

Recount, 29 Sep to 18 Oct 2026. No B row was read in source in the runs of 1–3 Oct, so 3 Oct added only new A rows. On 4 Oct, 2 B rows (jev-trader, Jev Trade) were read in source and moved to A, and 6 new A rows were added. On 5 Oct, 3 B rows (Jev for Chrome, computer-use-jev, jev-android) were read in source and moved to A, and 5 new A rows were added. On 6 Oct, 20 new A rows were added (18 coding-agent integrations, 2 guard rows) and no row changed grade. On 8 Oct, 4 B rows read in source at pinned commits moved to A (jev-reranker, Python, at v0.1.2; pg-jev, sqlite-jev and DocJev, whose failure cells were read for Jev inside SQL and document pipelines), and 2 new A rows were added (LanceDB `TypeSafeReranker`, pg-redact). On 10 Oct, 2 new A rows were added (jev-effort-router, jev-gateway) and no row changed grade. On 11 Oct, 3 new A rows were added (jev-triage, the jev-for-all launcher, hermes-jev-mana) and no row changed grade. On 18 Oct, 14 new A rows were added (3 Claude Code guards, 3 game and robot demos, 8 projects) and no row changed grade.
Grade	29 Sep	Added 3 Oct	3 Oct	Changed 4 Oct	4 Oct	Changed 5 Oct	5 Oct	Added 6 Oct	6 Oct	Changed 8 Oct	8 Oct	Added 10 Oct	10 Oct	Added 11 Oct	11 Oct	Added 18 Oct	18 Oct
A · call path read	15	+24	39	+6 new, +2 from B	47	+5 new, +3 from B	55	+20	75	+2 new, +4 from B	81	+2	83	+3	86	+14	100
B · project docs describe Jev use	21	0	21	−2 to A	19	−3 to A	16	0	16	−4 to A	12	0	12	0	12	0	12
C · scripted or mocked	1	0	1	0	1	0	1	0	1	0	1	0	1	0	1	0	1
Total rows	37	+24	61	+6	67	+5	72	+20	92	+2	94	+2	96	+3	99	+14	113
Independently verified in production	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0

What the grades mean

Grade	What the pool checked	What it does not show	Evidence label
A	Read the source file that sends the Jev request	That the project runs, has users, or works well	Documented
B	The project's own repository describes a Jev integration; the call file was not read in that run	That the code matches the description	Reported
C	A scripted or mocked example	Any live Jev call	Documented as a mock
P	The creator claims live use	Live use, until checked independently	Reported

Project ledger

Project and source	Grade	Form	What Jev does in it	Limits and status	Checked (2026)
QuantDinger · filter source · mocked tests · pinned at a5a9f4c	A	Trading platform (optional gate)	Optional Jev entry gate: a direct TypeSafe request with five Choices; code rejects only on a risk or execution block or a confident two-sided directional conflict.	Filter off by default; strategy orders default to a virtual account. On a Jev error, malformed answer, low confidence (below 0.55) or no key it tries an LLM, then allows the order: `fallback-LLM`, then `fail-open` (error path, no key). Exits, grid, DCA and martingale strategies and QuantDinger's own credit denial skip the gate. Order call sites traced 4 Oct: the gate runs before the order in strategy live orders, quick trade and the Alpaca route. No live Jev call, order or production use verified. Compared on trading and fraud gates.	28 Sep; re-read 4 Oct at `a5a9f4c`
Jevmail · classifier · schema · pinned at f6f20af	A	Gmail application	Read-only OAuth; AI SDK `typesafe-ai/jev` tray Choice, urgency Score and a personal-mail question.	Low confidence is a flag, not a review gate; an unexpected category maps to promotional. On a Jev error the email stays unclassified and is retried on the next run; repeated 429s pause the run and a 403 stops it. No-key `/preview` simulates without Jev. Live use, account access and cost unverified. Classifier unchanged since the 28 Sep audit (last commit 19 Sep). Compared on email and lead sorting.	28 Sep; re-read 3 Oct
Jev Comment Triage · plugin source · pinned at d848a5b	A	WordPress plugin	Holds eligible comments, then async spam/scam Noul and toxicity Score against site thresholds.	Provider errors keep comments held, then human review after three failures; with no provider configured WordPress's original decision stands, possibly publish. Default payload includes author identity. README timings are creator-reported. Compared on content moderation.	28 Sep; pinned 2 Oct
sqlite3-jev	A	SQLite C extension	Direct libcurl call path from SQL.	SQL input may leave the local database. Failure cells read at `2172600`: SQL and documents row.	25 Sep
browser-use/jev-ultrafast (`1231850`) · model.py	A	Browser agent demo	Default `jev-latest`; posts page and goal state with operation and target questions, then validates choices.	Fail-closed on a Jev error (25 s timeout; "no action executed"); acts on any confidence (no threshold); runs in the user's existing Chrome. README flight timing is maintainer-reported; no cloud-product operation shown. browser analysis	5 Oct
Canny · src/jev.ts	A	Agent-supervision tool	Advisory completion judgment beside a deterministic ledger.	Keyed mode sends a clipped diff and final message. Session traces and evaluations are author-reported; installations unknown.	25 Sep
killmyidea · typesafe.ts · evaluate.ts	A	Web application	Direct server-side Jev call with application scoring and public-deployment metadata.	Public host was inaccessible in the 23 Sep audit; operation and adoption unverified.	23 Sep
json-render Jev playground · compose.ts	A	Experimental playground	`typesafe-ai/jev` via Vercel AI Gateway; component membership and layout Choices.	Synthetic data; v0.21.0 released 18 Sep. No customer deployment shown. Graded A/B.	25 Sep
typesafe-router (TypeSafeAI community organisation)	A	Community-organisation repo (github.com/TypeSafeAI), not TypeSafe official	Server Choice call with a review fallback.	The TypeSafeAI organisation describes itself as "an UNOFFICIAL organization made for community repos" (created 18 Sep). No-key demo uses `mockCallJev`. An example, not a customer deployment. Graded A/C.	25 Sep; label corrected 28 Sep
DocJev	A	Document application	Extracts text, then category and split judgments.	Optional cloud OCR changes the data path. The author's 19 Sep 40-document comparison excludes OCR time. Failure cells read at `c7abe27`: SQL and documents row. Graded A on 8 Oct (call path read at the pinned commit).	8 Oct
pg-redact (`ee903ef`)	A	Next.js demo app writing to Postgres	A Choice per candidate span (PII type or "none") decides which spans are redacted before the message is stored.	fail-closed on a Jev error or no key (502, no insert); a missing span answer or a span below 0.5 is stored unredacted (fail-open; acts-anyway). Sends the whole pasted message. SQL and documents row	8 Oct
pg-jev	A	Postgres extension	SQL judgments.	Needs plpython3u and superuser; row data may leave the host. Author batch tests degrade at larger sizes. Failure cells read at `8d9598d`: SQL and documents row. Graded A on 8 Oct (call path read at the pinned commit).	8 Oct
sqlite-jev	A	SQLite extension	Batched `jev_rows` with a `max_rows` guard.	Calls TypeSafe, not a local model. Failure cells read at `1ac946c`: SQL and documents row. Graded A on 8 Oct (call path read at the pinned commit).	8 Oct
WordPress AI Provider for Jev	B	WordPress provider plugin	Server key and authenticated REST/PHP primitives.	Manual install; its fallback differs from the comment plugin.	25 Sep
jev-reranker (CLI)	B	Retrieval CLI	Reranks JSON candidates.	The answer still needs source verification.	25 Sep
jev-reranker (Python) · PyPI 0.1.2 (tag v0.1.2 = `d58594b`)	A	RAG library	Concurrent passage reranking: a Noul per passage, sorted, cut below the threshold (`relevance_rerank` 0.2), then `top_k`.	fail-closed on a Jev error (up to 8 retries, 180 s per request) and on a malformed answer; no key fails closed. No independent retrieval gain measured. Moved from B to A on 8 Oct 2026. reranking row	25 Sep; read in source 8 Oct
LanceDB `TypeSafeReranker` (in v0.40.0, `b1b080b`)	A	Vector-database reranker	A Noul per result ("whether the document is relevant"); sorts results by score, no cut.	fail-closed on a Jev error and on a malformed answer (errors propagate); no-threshold; no key not recorded. reranking row	8 Oct
openclaw-typesafe-ai	B	Agent tool	`typesafe_decide` calls the direct endpoint.	No installation or adoption evidence. A community project, not the vendor-published `@openclaw/typesafe` (added 6 Oct).	25 Sep
omp-jevens-classifier	B	Permission-gate prototype	Classifies commands; critical patterns, environment and call failures still prompt.	A judgment is not authorisation.	25 Sep
fast-jev-compaction	B	Developer tool	JevClient with `jev-latest`; keep/drop context judgments.	Adoption unknown; issue #52 describes possible over-dropping.	25 Sep
SemDecide	B	Released CLI	Typed commands, JSONL filtering and a provider-failure exit path.	No user's production pipeline verified.	25 Sep
jev-review	B	Local project	Staged code-review triage and dashboard.	No production status established. A different project from NiazMorshed2007/jev-review (added 6 Oct).	25 Sep
jev-usecases	B	Reference harness	Client and examples for Choice, Score and Noul.	README says it is not ready for unattended production use.	23 Sep
Jev Workbench	B	Local judgment service	HTTP and MCP functions.	No-key demo is simulated; the maintainer claims a keyed test. No idempotency or retry protection. Graded B/C.	25 Sep
Jev Logs	B	Observability tool	Vercel live mode.	Default is a fixed offline demo; its evidence note says the live test was skipped. Savings are modelled. Graded B/C.	25 Sep
AgentRun · Show HN	B	SDK	`@parcha/agentrun-jev`, `createJevRunner()`, optional live key route.	Default demos are scripted and need no key; live operation unverified. Graded B/C.	25 Sep
Jev Showcase	B	Unofficial showcase	Local and Docker examples with an optional live key.	No public operation verified. Graded B/C.	23 Sep
typesafe-playground tool-router demo (TypeSafeAI community organisation)	C	Community-organisation repo (github.com/TypeSafeAI), not TypeSafe official	Jev routing example.	Agent and tool execution are explicitly mocked. TypeSafe's official organisation is github.com/typesafe-ai.	23 Sep; label corrected 28 Sep

Claude Code, MCP and agent-guard integrations

Audited in source at pinned versions on 29 Sep 2026; the last 2 rows were added on 6 Oct 2026. These rows describe what the code does; none is a production product, and star counts are not adoption. The full comparison is on Claude Code and MCP: which server, hook or plugin fails open? Three Claude Code guards added on 18 Oct 2026 are in the 18 Oct table below. Documented

Integration and source	Grade	Form	What Jev does in it	Failure behaviour and status	Checked (2026)
jev-mcp (jkudish) · tool source at v0.11.0	A	MCP server (12 tools)	Choice, Score and Noul-backed judgment tools; TypeSafe direct, OpenRouter, Cloudflare, Vercel or a compatible endpoint.	Advisory: returns recommendations; the server never blocks. Malformed answers become "review", not "pass". Read at 0.11.0 (`53fe575`, 29 Sep 2026); checked at 0.13.0 (`5e0ca5c`, 2 Oct) and 0.14.1 (`86eae78`, 8 Oct 2026): failure path unchanged (error returned to the agent; 60 s). Renamed from "jev-mcp" (at least four unrelated repositories use that name) and regraded from B (25 Sep). No client installation verified. failure matrix	29 Sep
danna-zhou/jev-mcp	A	Claude Code hooks and MCP server	PreToolUse hook asks a risk Choice and a needs-human Noul about high-stakes Bash commands; Stop hook scores the final message.	Hooks fail open on every failure class (unreachable endpoint tested offline by the pool); MCP tools are advisory. Commit `436469a` (23 Sep). No users verified. failure matrix	29 Sep
jev-claude (brunopivetta88)	A	Claude Code hook plugin	Six Nouls and a severity Score about each proposed tool call, mapped to allow, warn, ask or block.	Shadow by default: nothing is enforced until the mode is changed; enforce mode fails open by default. 0.1.0 (25 Sep). No users verified. failure matrix	29 Sep
LangChain `AutoModeMiddleware` (`langchain-typesafe` 0.0.1a3, commit `4af7ab8`)	A	Agent middleware from a framework vendor, not TypeSafe	One "is_risky" Noul per call to a listed tool; blocks the call at 0.5 or above.	Fail-closed: any Jev error ends the agent run. Sends the last 30 messages in full. Sends no risk criteria unless the caller passes them. Experimental alpha (20 Sep). failure matrix Noted 7 Nov 2026: `langchain-typesafe` 0.0.2 (5 Nov 2026) is newer than the 0.0.1a3 read; not re-read. (R27-S25)	29 Sep
oh-my-claudecode Jev judgment points (audited v5.5.0, commit `9fd35ec`)	A	Jev layer inside a Claude Code plugin	Optional judgment points (keyword detection, task size, model routing, compaction and others) compared with its own heuristic.	As installed (default shadow): advisory. Active (opt-in): fallback-heuristic: its heuristic decides on any Jev failure. Nothing is sent unless `OMC_JEV` lists points. Rechecked at v5.6.0 (`e74b22c`), failure path unchanged (3 Oct 2026); default timeout 2,000 ms from v5.6.0, 250 ms up to v5.5.0. New in v5.6.0: `OMC_JEV=<point>:active` makes Jev decide (a warning is printed). Checked v5.6.2 (GitHub release, 6 Oct 2026; commit `454bae0`) on 8 Oct 2026: failure path unchanged (advisory as installed; fallback-heuristic when active; 2,000 ms). failure matrix	29 Sep; v5.6.0 3 Oct
Official TypeSafe plugin `typesafe@typesafe-ai`	A	Claude Code plugin (official TypeSafe)	None: a skill that tells the agent to read TypeSafe's docs.	Instructions only: makes no Jev call and is not a guard. v0.5.7 (12 Sep). Its skill links a migration guide that returns 404. failure matrix	29 Sep
madisonrickert/jev-permission-gate (`7349bc9`)	A	Claude Code plugin (enforcing)	Decides permission requests for Bash, WebFetch and WebSearch in auto mode.	fallback-built-in classifier on a Jev error, 1.5 s timeout, malformed answer, low confidence or no key. Enforcing by default. Secret-like or too-long input is not sent. Run-7 pin; HEAD unchanged on 6 Oct. Added to the ledger 6 Oct. No users verified. guard row	6 Oct
gulbaki/jev-llm-guard (`7c7b1b5`)	A	Guardrail library and CLI	Returns an allow, review or block analysis of a prompt.	Returns verdict; not a decision: the `jev-guard` CLI exits 0 whatever the verdict and exits 1 only on an error or input over 20,000 characters; the demo server enforces nothing. A CI step that runs it fails only on an error, not on "block". Run-7 pin; HEAD unchanged on 6 Oct. Added to the ledger 6 Oct. No users verified. guard row	6 Oct

Added 3 Oct 2026: 24 grade-A rows in five domains

All 24 are grade A: the pool read the code that calls Jev at the commit linked in each row. Documented No live operation, users or production use is verified for any of them; the pool installed and ran none. Each group's full comparison is on its analysis page. How each group was picked is stated there: routers, moderation and email projects by GitHub stars (a reading order, not a sample), ticket-triage projects from three GitHub searches, n8n nodes from npm searches. None of the groups is a survey.

What happens on a Jev error or timeout, across the 24 new rows (default settings)

fail-closed: stops without acting14
fail-open: the item passes unchecked4
fallback: a fixed lane, model or local choice4
no-decision: an error goes back to the caller, no action2

Support-ticket triage5 rows
- fail-closedtuhinmitra888
- fail-closedenderkus
- fail-closedgbesse
- fail-closedveermshah
- fallbackCyrilBaah: archive lane
Model routers5 rows
- fail-closedagent-router
- fallbackgargpratyush: current model
- fallback0xNatoshi: frontier model
- fallbackautojev: local selection
- no-decisionJevRouter
Content moderation5 rows
- fail-openJev-Moderation-Bot
- fail-opensoter
- fail-openjev_antispam_bot
- fail-openmastra-jev-moderation
- no-decisionjev-demo-moderator
Email and lead sorting4 rows
- fail-closedilyamk Gmail labeler
- fail-closedjev-mail-classifier
- fail-closedjev-mail
- fail-closedspam-leadgen-pipeline
n8n nodes5 rows, 1 official
- fail-closedOfficial TypeSafe node
- fail-closedDomMonte
- fail-closedn8n-nodes-jev
- fail-closedtaifoon-io
- fail-closedzampierid4p

Colours follow the shared failure legend (green fail-closed, red fail-open, blue fallback, gray no-decision; changed 6 Oct 2026, when this page used blue for fail-closed and amber for fallback). Totals: fail-closed 14, fail-open 4 (all four are moderation bots or guards), fallback 4, no-decision 2; 24 in all. For n8n nodes, fail-closed means the node stops the workflow while Continue On Fail is off. The counts describe these 24 rows only, not how Jev projects are usually built. The colours show where an item goes, not whether that is right: in moderation, fail-open leaves a message up; in triage, fail-closed leaves a ticket unrouted. Documented from source read 1–3 Oct 2026.

Support-ticket triage · 5 rowsfail-closed 4 · fallback 1 · checked 1 Oct

Full comparison: Support-ticket triage: routing tickets to queues

Project (pinned commit)	Form	What Jev does in it	On a Jev error	Limits and status	Checked · analysis
CyrilBaah/jev-triage-desk (`6391317`)	Runnable demo (Next.js, seed tickets)	Choice department, Score urgency, Nouls escalate and spam pick a ticket lane	fallback-archive An errored ticket is put in the archive lane	No threshold on the department choice; no tests; demo data	1 Oct · triage analysis
tuhinmitra888/ai-jev-ticket-triage (`3e8b83a`)	Runnable demo (CLI and web; official JS SDK)	Category Choice with `other`, urgency and frustration Scores, four Nouls	fail-closed HTTP 502, ticket not routed	Low confidence goes to human triage; no live use shown	1 Oct · triage analysis
enderkus/zammad-jev-dispatcher (`f999eca`)	Helpdesk integration (Zammad webhook)	One Choice over Zammad groups; changes the ticket's group	fail-closed HTTP 502, ticket stays "Unclassified" with no note	No timeout set; no tests	1 Oct · triage analysis
gbesse/zammad-jev-triage (`5d2a0a7`)	Helpdesk integration (Zammad webhook)	One Choice; pinned `jev-1.13.0`; validates answers	fail-closed HTTP 503, ticket unchanged, Zammad retries	20 s timeout; tests and CI	1 Oct · triage analysis
veermshah/jev-support-desk (`aef8fa0`)	CLI (official Python SDK)	Queue and knowledge-base Choices, two Scores, three Nouls; labels and suggestions only	fail-closed The whole run exits	Created 30 Sep; tests and CI	1 Oct · triage analysis

Model routers · 5 rowsfail-closed 1 · fallback 3 · no-decision 1 · checked 2 Oct

Full comparison: Jev Router or your own model router?

Project (pinned commit)	Form	What Jev does in it	On a Jev error	Limits and status	Checked · analysis
gargpratyush/jev-router (`38da6b8`)	Claude Code proxy (model router)	Choice over Claude tiers plus three Scores	fallback-current model No change of model	1.5 s per attempt, 3 s deadline; tests	2 Oct · routing analysis
BillionsBobby/JevRouter (`e11084c`)	Routing library (models, tools, subagents)	Choice over capability manifests	no-decision Returns "no decision" to the caller	15–20 s timeout; tests	2 Oct · routing analysis
0xNatoshi/jev-codex-router (`8701ef7`)	Codex router	Picks tier, thinking depth and route lease	fallback-frontier model `gpt-6-astra`, also with no key	Sends only the current ask, clipped to 400 characters	2 Oct · routing analysis
nidhi-singh02/agent-router (`fb24d06`)	CLI (picks agent tool, model, effort)	Six Jev questions, then a ranker	fail-closed Nothing launched	Refuses to send text that looks like a credential; tests	2 Oct · routing analysis
thinkany-ai/autojev (`d5d5731`)	Desktop gateway (Tauri)	Choice over candidate models; no prompt text sent	fallback-local selection	2–8 s timeout; some tests	2 Oct · routing analysis

Content moderation · 5 rowsfail-open 4 · no-decision 1 · checked 2 Oct

Full comparison: Content moderation: what projects do when Jev fails

Project (pinned commit)	Form	What Jev does in it	On a Jev error	Limits and status	Checked · analysis
brainstormity/Jev-Moderation-Bot (`325ea7f`)	Discord moderation bot	Choice (legitimate / spam / scam link) plus a ban-review Noul; deletes from 0.70	fail-open Message left up	Sends author ID and account age; no step restores a deleted message	2 Oct · moderation analysis
frolleks/soter (`756579a`)	Discord moderation bot	Hate-speech Noul plus spam-level Choice	fail-open "a scan error never blocks chat"	Sends account and join dates and recent messages	2 Oct · moderation analysis
backmeupplz/jev_antispam_bot (`0b07e88`)	Telegram anti-spam bot	29 Nouls; deletes from 0.81	fail-open Message kept	Sends sender profile and recent messages; no reversal path	2 Oct · moderation analysis
CodeAlive-AI/mastra-jev-moderation (`8735ea8`)	Input guard for Mastra agents	"Must this be blocked?" Noul aborts the turn at 0.7	fail-open Message let through (5 s timeout, circuit breaker)	No tests	2 Oct · moderation analysis
Zafer-Liu/jev-demo-moderator (`cc47b25`)	Runnable demo (API and console)	Noul spam, Score toxicity, Choice action	no-decision Item returned with an error, no action	No publishing system; no tests	2 Oct · moderation analysis

Email and lead sorting · 4 rowsfail-closed 4 · checked 3 Oct

Full comparison: Email and lead sorting with Jev (new on 3 Oct). Jevmail, already in the ledger above, is compared there too.

Project (pinned commit)	Form	What Jev does in it	On a Jev error	Limits and status	Checked · analysis
ilyamk/jev-gmail-ai-spam-filter-and-labeling (`ade42bf`)	Gmail labeler (Apps Script; OpenRouter)	Choice over up to 12 label rules; labels, and archives spam at 0.93 or above	fail-closed Job pauses, no Gmail change	Preview by default; OAuth scope not declared	3 Oct · email analysis
parth-kp/jev-mail-classifier (`db96194`)	IMAP CLI	One Noul per category; tags, moves, flags, webhooks	fail-closed Run stops	Full IMAP access; live by default; body sent uncut	3 Oct · email analysis
vynnlee/jev-mail (`61e346f`)	Gmail worker on a timer (Apps Script)	Two Nouls plus a bucket Choice; labels, stars, archives	fail-closed Retried later with backoff, Gmail untouched	Requests full `mail.google.com` scope; unsure mail gets a "Review" label	3 Oct · email analysis
automatewithuday/spam-leadgen-pipeline (`78a7592`)	Lead-qualification script (Gmail spam folder to Supabase)	Two Choices plus a `legit` Noul mark senders as qualified leads	fail-closed Script ends	Read-only Gmail; created 2 Oct	3 Oct · email analysis

n8n nodes · 5 rows, 1 officialfail-closed 5 by default · checked 3 Oct

Full comparison: Jev n8n nodes: who publishes them and what they do when Jev fails (new on 3 Oct). Only the scoped @typesafe-ai/n8n-nodes-typesafe-ai is TypeSafe's; the unscoped n8n-nodes-typesafe-ai is a third party's.

Node (pinned commit)	Form	What Jev does in it	On a Jev error	Limits and status	Checked · analysis
Official `@typesafe-ai/n8n-nodes-typesafe-ai` 0.9.0 (`c12537b`)	n8n node (official TypeSafe; verified on n8n)	Evaluate and Route (Choice, Noul, Score)	fail-closed by default; with Continue On Fail, the Fallback or Uncertain output if configured, otherwise the first route	npm provenance to the typesafe-ai tag 0.9.0; 5 s timeout	3 Oct · n8n analysis
DomMonte/n8n-nodes-typesafe-ai (`ff2afca`; npm `n8n-nodes-typesafe-ai`)	n8n community node (third party; same name as the official repository)	askNoul, askChoice, askScore, evaluate	fail-closed by default	Not TypeSafe's; unverified on n8n	3 Oct · n8n analysis
vibe-with-me-tools/n8n-nodes-jev (`ddbfe85`)	n8n community node	ask, route	fail-closed by default; with Continue On Fail, the first route unless a low-confidence output is on	60 s timeout; unverified on n8n	3 Oct · n8n analysis
taifoon-io/n8n-nodes-typesafe (`cdcc0f3`)	n8n community node	Ask, Translate; pass / fail / review outputs	fail-closed by default; with Continue On Fail, `fallback-review`	Sends the whole item by default; unverified on n8n	3 Oct · n8n analysis
zampierid4p/n8n-nodes-typesafe (`549a545`)	n8n community node	ask (per-question branches)	fail-closed by default; with Continue On Fail, question 1's branch	Unverified on n8n	3 Oct · n8n analysis

Added 4 Oct 2026: 6 new grade-A rows and 2 moved from B, in two domains

All 8 are grade A: the pool read the code that calls Jev at the commit linked in each row, on 4 Oct 2026. Documented Two were already in the ledger as grade B (jev-trader and Jev Trade) and were read in source for the first time. The pool ran nothing: no bot, no order, no payment, no eval and no Jev call. No live operation, users or production use is verified for any of them. Trading and fraud projects were picked from projects named in earlier pool research, two GitHub searches and the 24-hour scan of 4 Oct; judging tools from tools named in earlier pool research and two projects found in that scan. Neither group is a survey. QuantDinger, also re-read on 4 Oct, stays in the ledger above.

What happens on a Jev error or timeout, across the 8 rows added or moved to A on 4 Oct (default settings)

fail-closed: stops without acting5
fail-open: the item passes unchecked0
fallback: a rules-only path1
no-decision: the item is left unscored2

Trading and fraud gates6 rows
- fail-closedjev-trader
- fail-closedJev Trade
- fallbackbtc-hft-jev: rules-only
- fail-closedjev-fraud-shield
- fail-closedfraud-jev
- fail-closedexpense-policy gate
Judging tools2 rows
- no-decisiondid-they-answer
- no-decisioncompetitor-hunter

Totals: fail-closed 5, fallback 1, no-decision 2, fail-open 0; 8 in all. For jev-trader and Jev Trade, fail-closed means no new order; the order already resting on the book is not cancelled. btc-hft-jev's rules-only fallback can still quote. None of the 3 fraud or payment gates has a payment call site. With QuantDinger (an LLM fallback, then the order is allowed), 2 of the 4 trading bots can still place a new order on a Jev error. The counts describe these rows only. Documented from source read 4 Oct 2026.

Trading and fraud gates · 6 rowsfail-closed 5 · fallback 1 · checked 4 Oct

Full comparison: Jev trading bot: does a Jev failure still place the order? (new on 4 Oct). All 4 trading bots there, QuantDinger included, start in mock, paper, testnet or virtual mode, and all 4 have a path that trades without Jev. No performance figures and no trading advice.

Project (pinned commit)	Form	What Jev does in it	On a Jev error	Limits and status	Checked · analysis
jarrodwatts/jev-trader (`b587759`); B until 4 Oct	Trading bot (Monad, Kuru)	With `MODEL=jev`, picks buy or sell; any answer other than "sell" becomes a buy	fail-closed No new order; the resting order is not cancelled	Default `MODEL=mock`; dry run if no wallet key is set or `DRY_RUN=true`. "Mock" means no Jev, not no trading: the mock model still sends real orders when a wallet key is set. No-key path not read (inside `@ai-sdk/typesafe-ai`). The public Railway endpoint was a dry-run mock (25 Sep). No live Jev trade verified	4 Oct · trading analysis
aowang-ai/jev-trade (Jev Trade, `a3f2f83`); B until 4 Oct	Trading desk (Hyperliquid); the creator claims live fills	With `MODEL=jev` and a key, decides the next quote from market data and the open position; 4,000 ms deadline	fail-closed No new order; the previous quote is not pulled	Defaults `MODEL=mock` and `HL_TESTNET=true`; dry run without a wallet key. "Mock" means no Jev, not no trading: the mock model still sends real orders when a wallet key is set. With Jev selected and no key it refuses to start. The creator's live-fill claims are Reported, not checked; the public desk was Offline with blank positions on 27 Sep. Graded A/P	4 Oct · trading analysis
taran1610/btc-hft-jev (`96f3541`)	Trading bot (Alpaca); created 3 Oct	Judges market conditions for a quoting loop; below 0.50 confidence it trades at half size, and a low quote-environment score stands it down	fallback-rules-only Reuses a judgment up to 60 s old, then a rules-only path that can still quote	Paper by default; live needs `--live`, an environment flag and a typed phrase. No key: mock and dry run, no orders. A malformed answer stops the run and cancels resting orders. Dashboard pause and 5% drawdown stop. No live order verified	4 Oct · trading analysis
jonny5isalive5/jev-fraud-shield (`eeb4fc6`)	Runnable demo (synthetic transactions)	Judges a transaction, with cardholder free text cut to 200 characters; unsure results go to step-up	fail-closed Step-up (2 retries, no timeout in code)	Mock without a key (local mock fed simulated signals). 0 flags approve and 3 or more flags decline without Jev. No payment call site: returns a result object	4 Oct · trading analysis
frankied003/fraud-jev (`aa33b97`)	Runnable demo (Next.js; fictional payments)	Judges a payment from payer, beneficiary and account data; allow or review at fixed cuts	fail-closed Review, HTTP 500	Key pasted in the UI; no key gives HTTP 401. A non-numeric answer becomes allow (fail-open). Below the cuts it allows. No payment call site: JSON to the UI	4 Oct · trading analysis
ElliotOne/nl-jev-decision-gates-expense-policy (expense-policy gate, `161c33b`)	Expense-approval gate (CLI); created 3 Oct	Judges expense claims, with full receipt text, against a written policy; unsure claims go to manual review	fail-closed The batch stops before any output (30 s, no retries)	Live Jev by default; no key: refuses to run. Fixed rule rejects skip Jev; nothing is approved without Jev. No reimbursement call site: writes a decisions file	4 Oct · trading analysis

Judging tools · 2 rowsno-decision 2 · checked 4 Oct

Full comparison: Jev as a judge: accept when confident, escalate when unsure (new on 4 Oct). DeepEval JevEval and openlayer jevals are compared there too; they are library features and are not counted here.

Project (pinned commit)	Form	What Jev does in it	On a Jev error	Limits and status	Checked · analysis
Kandarp-Joshi-007/did-they-answer (`e0ee60c`)	Daily judging job plus site; created 4 Oct	Judges Irish parliamentary written answers: Choices for clarity and technique, a Noul for "referred"	no-decision The item is not scored and is retried on the next daily run	Sends the full question and reply. A malformed answer stops the run (traced, not run). No key: stops. No unit tests	4 Oct · judge analysis
charlesdove977/competitor-hunter (`40613ea`)	Local app (competitors' social posts); created 3 Oct	Choices for topic and hook type, Scores for audience fit and proof, Nouls for replicable, sponsored and news-driven posts	no-decision The item is left out of the ranking; the run stops if more than 5% of items error	Loads the Jev caller the user has installed, so that part is not pinned. Unsure items are flagged but still ranked. Needs an OpenRouter key. No tests	4 Oct · judge analysis

Added 5 Oct 2026: 5 new grade-A rows and 3 moved from B, in browser and computer use

All 8 are grade A: the pool read the code that calls Jev at the commit linked in each row, on 5 Oct 2026. Documented Three were already in the ledger as grade B (Jev for Chrome, computer-use-jev and jev-android) and were read in source for the first time. browser-use/jev-ultrafast was re-read at a new commit and stays in the ledger above. The pool ran nothing: no browser or computer-use agent, no extension install and no Jev call. No live operation, users or production use is verified for any of them. Projects were picked from leads in earlier pool research, three GitHub searches, Hacker News and the 24-hour scan of 5 Oct; this is not a survey. The trycua/cua example uses a mock provider by default; second-look gives advice and never acts.

What the 7 browser and computer-use agents read on 5 Oct do (default settings; jev-ultrafast included, second-look and trycua/cua not)

Still clicks, types or taps on a Jev error0/7
Acts on any confidence (no threshold)5/7
Acts in the user's own browser, phone or desktop session5/7
Asks a person before send, buy or delete1/7

Browser and computer use8 rows added or moved; on a Jev error
- fail-closedJev for Chrome
- fail-closedcomputer-use-jev
- fail-closedjev-android
- fail-closedironbee-express
- fail-closedjev-agent-browser
- fail-closedUlka
- fallbacksecond-look: heuristic verdict
- mock by defaulttrycua/cua

Counts over the 7 agents that click, type or tap (jev-ultrafast, ironbee-express, jev-agent-browser, jev-android, computer-use-jev, Jev for Chrome, Ulka). On a malformed answer, 0 of 7 act directly; jev-agent-browser can click a fallback target. With no key, 6 of 7 refuse; ironbee-express replays a cached recording unjudged. computer-use-jev blocks mutating actions below 0.5 confidence and jev-android stops below 0.65. The counts describe these rows only. Full table and risks: Jev browser use and computer use. Documented from source read 5 Oct 2026.

Browser and computer use · 8 rowsfail-closed 6 · fallback 1 · mock 1 · checked 5 Oct

Full comparison: Jev browser use and computer use (TypeSafe AI) (new on 5 Oct). Speed and benchmark claims for these agents are the authors' own (Reported).

Project (pinned commit)	Form	What Jev does in it	On a Jev error	Limits and status	Checked · analysis
chy4pro/jev-for-chrome (`d5c24de`); Jev for Chrome; B until 5 Oct	Browser extension	Chooses the next operation and a target per operation; asks whether the goal is done or stuck	fail-closed 3 retries, then "no action executed"; no request timeout	Acts in the user's own tab and logins. Sends visible text up to 6,000 characters and field values; password, file and hidden inputs excluded. No threshold for actions. No key: refuses. No guard before irreversible actions found. A store listing shows distribution, not task success or active users	5 Oct · browser analysis
paulsmith/computer-use-jev (`ff0ad8b`); B until 5 Oct	macOS prototype	Chooses the next action, target and shortcut through the Accessibility API	fail-closed 60 s timeout; the loop returns an error	Acts in the user's own macOS session. Below 0.5 confidence a mutating action is not executed. Passthrough mode runs without Jev or a key. Whether `[secure]` field values are sent: not recorded. Fake-worker tests do not show task success	5 Oct · browser analysis
dougsong/jev-android (`be8364a`); B until 5 Oct	Android SDK and sample	Chooses operation, target and text on the user's device	fail-closed 25 s call timeout; task deadline 120 s	Acts on the user's own device and apps. Below 0.65 it replans up to 2 times, then stops (BLOCKED). Password nodes skipped. The action gate allows every action by default. DONE is unverified without a host OutcomeVerifier; API unstable	5 Oct · browser analysis
ironbee-ai/ironbee-express (`05079b4`)	Browser agent CLI (Elastic License 2.0)	Chooses the next operation, target, tab, key and text	fail-closed 25 s timeout, 3 tries, then "no action executed"	Fresh headless profile with no cookies by default. Secrets masked. No threshold. No key: a cached recording replays unjudged (counted fail-open, conditional); without one it refuses. No guard before irreversible actions found	5 Oct · browser analysis
forvela/jev-agent-browser (`eb35251`); npm 0.2.1	Browser agent CLI	Chooses operation, target and input; judges goal reached or stuck	fail-closed 30 s timeout; the loop ends with an error	Own browser session by default. Sends up to 18,000 characters of snapshot text. A malformed answer can fall back to a target picked by word overlap, then act. No threshold for actions. The npm tarball was not compared to the repository	5 Oct · browser analysis
razaanstha/ulka (`bb04433`); Ulka	Chrome extension	Chooses operation and target through Vercel AI Gateway (`typesafe-ai/jev`)	fail-closed retries once by another route to the same model, then stops; no timeout found	Acts in the user's own Chrome. Sends field values up to 4,000 characters; passwords excluded. No threshold. Send, submit, buy, delete and similar actions need approval from a person in the side panel; with no approver they are blocked	5 Oct · browser analysis
KaushikSiva/second-look (`0c05d16`); created 4 Oct	Shopping-advice extension (recommendation only)	One shopping verdict: buy now, wait, skip or buy an alternative	fallback-heuristic verdict 15 s timeout; shows "Good deal — buy it" unless a fake discount was detected, with no confidence	Never acts; a person decides every purchase. Sends derived product signals, not form values. No key: same heuristic verdict. Whether the card says Jev was not consulted: not recorded	5 Oct · browser analysis
trycua/cua `jev-use` example (`ba0033a`)	Runnable demo (cua driver)	Chooses a candidate action ID	mock by default With a key and the live provider, the call is not wrapped, so an error stops the run (fail-closed)	The default provider is a mock, so as shipped it does not call Jev; listed for its call path, not counted on the browser page. Isolated new browser profile. No-key behaviour: not recorded	5 Oct · browser analysis

Added 6 Oct 2026: 20 new grade-A rows, 18 coding-agent integrations and 2 guard rows

All 20 are grade A: the pool read the code that calls Jev at the commit linked in each row, on 6 Oct 2026. Documented The 18 coding-agent integrations (Codex, OpenCode, Cursor, Hermes Agent, OpenClaw) are below. The 2 guard rows, madisonrickert/jev-permission-gate and gulbaki/jev-llm-guard, are in the Claude Code, MCP and agent-guard table above; both were read at their run-7 commits, unchanged on 6 Oct. The pool ran nothing: no agent was run and no plugin, hook, skill or MCP server was installed or started; no Jev call. Two npm packages were downloaded, read and deleted without being run. No live operation, users or production use is verified for any of them. Selection: the rows read for build/coding-agents plus the two guard rows; not a survey of coding-agent plugins.

One of the 18 is published by a host vendor: @openclaw/typesafe, by OpenClaw. None comes from TypeSafe. Two existing B rows have similar names but are different projects, and both stay: "jev-review" (devagrawal09) is not NiazMorshed2007/jev-review, and "openclaw-typesafe-ai" (Olli0103) is a community project, not the vendor-published @openclaw/typesafe.

Added 10 Oct 2026: 2 more grade-A coding-agent rows, jev-effort-router (Hermes Agent plugin, 562dc2a) and jev-gateway (npm 0.5.1, gitHead 37ff982), both in the table of rows that act on Jev's answer. Their code was read on 10 Oct 2026; nothing was installed or run and no Jev call was made. Documented

Added 11 Oct 2026: 3 more grade-A coding-agent rows, all in the table of rows that act on Jev's answer: jev-triage (OpenClaw plugin, 6360e21; host path read at OpenClaw aa6008a), the jev-for-all start launcher (npm 0.7.1, gitHead 74f62e7) and hermes-jev-mana (Hermes Agent core patch, 3daf639). Their code was read on 11 Oct 2026, 05:17–05:27 UTC; nothing was installed or run and no Jev call was made. Hoppin-Sorter/jev-router, also read on 11 Oct, is not settled on build/coding-agents and is not a ledger row. Documented

What happens on a Jev error, across the 13 coding-agent integrations that act on Jev's answer (13 of 24 read in source; recounted 11 Oct 2026)

Act on Jev's answer13 of 24 rows
- fail-closedjev-kit: held for the host's prompt
- fail-closedcodex-jev-approval-hook: held for Codex's approval
- fallbackrh-guard: lexical detectors
- fallbackopencode-tool-gate: OpenCode permission rules
- fallbackopencode-osuki-agent: default tier
- fail-openhermes-jev jev-judge (enforce mode)
- fallbackhermes-jev-approvals: main agent model, which can approve; held for a person on other HTTP errors; with no main model configured, it asks you (off after install)
- fail-closedhermes-jev-curator: background delete or patch blocked (observe by default)
- fallbackjev-effort-router: configured model (added 10 Oct)
- fallbackjev-gateway: the agent's own model picks the tool and the call runs (added 10 Oct)
- fallbackjev-triage: main model (added 11 Oct)
- fallbackjev-for-all launcher: default model, standard tier (added 11 Oct)
- fallbackhermes-jev-mana: original Hermes branch, configured model (added 11 Oct; inert after the documented install)
Read, not settled1 of 24 rows
- 1 of 1Hoppin-Sorter/jev-router
Not counted10 of 24 rows
- 8 of 10return Jev's answer to the agent
- 2 of 10advisory hints

On a Jev error the 13: fail-closed 3, fallback 9, fail-open 1. Does the tool call still run on a Jev error? Yes in 4 of 13 (1 fail-open, 3 fallback: OpenCode's permission rules, the default tier, the agent's own model), conditionally in 2 (both fallback), no in 3 (fail-closed). The other 4 are model-choice or routing rows (fallback): the turn runs on another model. 13 + 10 + 1 = 24. With no key, 5 hold the call or do not start, 4 let it run, 3 fall back and 1 is not recorded (hermes-jev-mana). 3 of the 13 do nothing until you switch them on (jev-approvals is off after install; jev-curator observes by default; hermes-jev-mana is inert after the documented install because admit_unlabeled defaults to false), and jev-judge starts in shadow mode. jev-effort-router is off until enabled in Hermes, and with Hermes defaults (model "") it is inert. The counts describe these rows only and match How to use Jev with Codex, OpenCode, Cursor and Hermes Agent, where the full table, checklist and sources are. Documented from source read 6, 7, 8, 10 and 11 Oct 2026; recounted 11 Oct 2026. Change, 18 Oct 2026: the outcome grouping "3 hold, 5 run, 5 fall back" is replaced by the grouping above, which matches the failure terms on When Jev fails; the 3 / 9 / 1 counts are unchanged. Changed 11 Oct 2026: was 10 of 22 (3 hold, 5 run, 1 may ask the main model, 1 keeps its configured model)

Coding agents that act on Jev's answer · 13 rowsfail-closed 3 · fallback 9 · fail-open 1 · checked 6–7 Oct; 2 added 10 Oct; 3 added 11 Oct

Project (pinned commit)	Form	What Jev decides	On a Jev failure	Limits and status	Checked · analysis
keiffff/jev-kit (`3d74b05`) `jev-agent-review`	Codex and Claude Code permission hook	Approves a permission request, or leaves it to the host's own prompt; it never denies	fail-closed held for the host's own prompt on no key, timeout, HTTP error or malformed answer; below its thresholds: held	Enforcing (approve-only) once configured. Requests with secret-like values or sensitive paths are held without a Jev call. Only calls the host sends as permission requests reach it	6 Oct · coding-agents row
bloudhood/codex-jev-approval-hook (`b6242c9`)	Codex permission hook (PowerShell)	Allow, ask you or deny for the approval request	fail-closed held for Codex's own approval; timeout 5 s by default; below its thresholds: held	Only Bash, apply_patch and MCP tools are matched. The endpoint and model are set by the user; no TypeSafe URL is fixed in the code. Stops deciding after repeated denials	6 Oct · coding-agents row
24601/rh-guard (`c2e682e`)	Codex and Cursor pre-tool hook	Jev labels add risk evidence that can move the verdict to block	fallback-lexical detectors on no key, 1.2 s timeout, HTTP error or malformed answer; low confidence: acts-anyway (steer message only)	Mode shadow, review or enforce; the default is enforce (checked 7 Oct), so a default install can block. On Codex, MCP tools are not matched. Redaction not read	6 Oct · coding-agents row
justmytwospence/opencode-tool-gate (`bdf1ef8`); 0.1.0	OpenCode plugin	Allows or holds edits, writes, patches, non-read-only shell commands and MCP calls	fallback-opencode permission rules on no key, 3 s timeout, HTTP error or malformed answer; below its thresholds: acts-anyway	Enforcing by default. Tool arguments up to 6,000 characters are sent, so content being written can leave the machine. OpenCode's default rules (v1.18.35, checked 7 Oct) allow every tool, so the fallback lets the call run. Created 6 Oct	6 Oct · coding-agents row
`@osuki-dev/opencode-osuki-agent` 0.2.7 (npm; tag `v0.2.7` = `c96ce03`)	OpenCode router plugin	Picks the subagent tier and a shortlist of tools	fallback-default tier on no key, 10 s timeout, HTTP error, malformed answer or low confidence	Jev is opt-in: the default decision provider is not TypeSafe. File contents in tool results can be sent (regex redaction only). Tarball not matched to the tag file by file. Checked at 0.2.8 (tag `v0.2.8` = `0ae9e88`): failure path unchanged; 0.2.8 tarball not matched to its tag file by file either	6 Oct; 0.2.8 checked 7 Oct · coding-agents row
DoGMaTiiC/hermes-jev (`5dddbea`); `jev-judge` and `jev-skill-router`	Hermes Agent plugins	`jev-judge` asks for confirmation of terminal, write and patch calls on risk signals; `jev-skill-router` adds a skill hint (advisory)	fail-open in enforce mode on no key, 3 s timeout, HTTP error or malformed answer; below its thresholds: acts-anyway	Shadow by default (logs only). Arguments are redacted and cut to 600 characters each. The skill router is listed with the judge, not counted separately	6 Oct · coding-agents row
anpicasso/hermes-jev-approvals (`28be98a`) (catalog commit); `jev-approvals` 0.3.0	Hermes Agent approval provider	Hermes approval provider: on APPROVE, Hermes core runs the flagged command without a person; DENY or ESCALATE goes to a person or is blocked	fallback-main agent model on a Jev timeout, connection error, 429 or 402 when a main provider is set: Hermes core hands the question to the main agent model, which can approve the command (corrected 8 Oct 2026; `auxiliary_client.py` L7610–7701 at `f97608f`). With no main model configured, it asks you (Hermes default model `""`, R13-S217). Held (escalates to a person) on other HTTP errors, no key or a malformed answer (core `v2026.9.24`). Below confidence 0.55: held	Off after install: Hermes core's default approval mode is `manual`; it acts only with `approvals.mode: smart`. Counted on the coding-agents page from 7 Oct	7 Oct; error cell corrected 8 Oct · coding-agents row
anpicasso/hermes-jev-curator (`4e8626c`); `jev-curator` 0.1.0	Hermes Agent skill curator	Hermes skill curator: in guard mode, plans built from Jev's judgments decide which background skill deletes or patches may go ahead	fail-closed on timeout or HTTP error: no judgment, so an uncovered background delete or patch is blocked. No key and malformed answer: fail-closed (no judgment, so no plan); below confidence 0.85: held (filled 8 Oct 2026 at `4e8626c`)	Observe by default: blocks nothing until the mode is set to `guard` or `apply`. Counted on the coding-agents page from 7 Oct. A separate repository from jev-approvals	7 Oct; cells 8 Oct · coding-agents row
AlphaPerseii3000/jev-effort-router (`562dc2a`); v0.2.2	Hermes Agent plugin (`llm_request` middleware)	Picks the Ollama:Cloud model and reasoning effort once per user turn; the plugin rewrites `model` and `reasoning_effort` in the outgoing request	fallback-configured model on a 2.0 s timeout, HTTP error, malformed whole answer or no key: the request is left unchanged. Below confidence 0.5 or an unknown choice: fallback-default model, the code rewrites the request to `deepseek-v4.1-flash` at medium effort	Decides model and effort only, not tools or permissions. Routing is on in the plugin's own config, but the plugin is off until enabled in Hermes; with Hermes defaults (model `""`) it is inert, because it acts only when the provider is Ollama:Cloud (R13-S217; wording corrected 11 Oct 2026). Nothing was installed or run; no live operation, users or production use verified	10 Oct; wording 11 Oct · coding-agents row
`jev-gateway` 0.5.1 (npm; gitHead `37ff982`)	Local proxy for Codex, OpenCode, Claude Code, Kilo, Gemini and Devin (community npm package)	Which tool the agent calls: the gateway forces `tool_choice` or builds the tool call	fallback-agent's own model on a 4,000 ms timeout (plus 1 retry), HTTP error or malformed answer, and below confidence 0.7: the request passes through, the agent's model picks the tool and the call runs. No key: fail-closed (the gateway does not start)	Tarball read at 0.5.1; nothing was installed or run. No live operation, users or production use verified	10 Oct · coding-agents row
RanyAlbegWein/jev-triage (`6360e21`); 1.0.3; host OpenClaw `aa6008a`	OpenClaw plugin (`before_dispatch` hook)	At 0.9 or above, "thanks" gets a canned reply and "spam" ends the turn silently; the host then skips the main model, with no person involved	fallback-main model on no key, a 5,000 ms timeout, HTTP error or malformed answer: the message goes to the main model as if the plugin were absent. Below 0.9: fallback-main model	On at startup; acts only where Jev is the agent's decision model (README). Routing, not a tool gate. Nothing was installed or run; no live operation, users or production use verified	11 Oct · coding-agents row
`jev-for-all` 0.7.1 (npm; gitHead `74f62e7`), `start` launcher	Launcher for Claude Code and OpenCode (plus an OpenCode plugin)	Picks the model tier and effort, then launches `claude` or `opencode` with that model, with no person involved	fallback-default model (standard tier, claude sonnet, medium effort) on no key, a 3,000 ms timeout, HTTP error, malformed answer or below threshold	Counted on the `start` launcher; the OpenCode plugin narrows the tool list and is not settled; the Claude Code hook was not read. Nothing was installed or run	11 Oct · coding-agents row
abehrman/hermes-jev-mana (`3daf639`); core patch targeting Hermes `e97923c3`	Hermes Agent plugin and core patch	When admitted: model lane and effort, a direct answer without a provider call, an executor run without a person, and tool calls built from a chosen plan	fallback-original Hermes branch (configured model) on a timeout (1.0 s per call / 5.0 s per turn + 0.25 s grace), HTTP error, malformed answer or below threshold; no key: `not recorded`	Inert after the documented install: every Jev call is refused because admit_unlabeled defaults to false; acts only if the operator sets admit_unlabeled: true. Nothing was installed or run	11 Oct · coding-agents row

Coding agents that return Jev's answer or advise · 10 rowsreturned to the agent 8 · advisory 2 · checked 6 Oct

These rows read Jev's answer in source but do not themselves stop or allow a tool call, so they are not counted on build/coding-agents. Hermes Agent catalog entries that install a commit older than HEAD (here and the counted jev-approvals row) link the catalog commit (Hermes pins).

Project (pinned commit)	Form	What Jev decides	On a Jev failure	Limits and status	Checked · analysis
`@openclaw/typesafe` 2026.9.8 (npm) · source in openclaw/openclaw (`aa6008a`)	OpenClaw decision provider; vendor-published (host vendor OpenClaw; SLSA provenance from the `openclaw/openclaw` repository)	Returns Jev's answer to the agent through OpenClaw's decision tool; it gates nothing	no-decision on no key, 30 s timeout or HTTP error: the agent sees "unavailable"	Disabled by default. Not a TypeSafe product. The source `package.json` at that commit says 2026.9.7; the published version is 2026.9.8. OpenClaw core not read. Not the B row openclaw-typesafe-ai above, which is a separate community project	6 Oct · coding-agents row
DECRUX9812/typesafe-skill-router (`e6cdac2`) (catalog commit)	Hermes Agent skill hint	Adds a skill-relevance hint; the agent's model still picks the skill (advisory)	no-decision on no key, 10 s timeout or HTTP error: no hint is added	Off by default. HEAD is a later commit than the catalog installs	6 Oct · coding-agents row
ourines/hermes-jev (`28a4ea3`) (catalog commit); `jev` 0.1.2	Hermes Agent tool	Returns Jev's answer to the agent (`jev_evaluate` tool)	No key: `not recorded`. HTTP error: the error is returned to the agent; 30 s timeout	HEAD v0.2.0 (not in the catalog) adds model routing, off by default; see model routing	6 Oct · coding-agents row
ajensenwaud/hermes-jev-plugin (`b3d29f7`); `jev-typesafe` 0.1.0	Hermes Agent tools (4)	Returns Jev's answer to the agent	no-decision on no key or HTTP error; error text goes to the agent; 30 s timeout	Tool only; gates nothing	6 Oct · coding-agents row
ydmw74/jev-skill-router (`76ee09e`); `jev-skill-router-mcp`	Hermes Agent MCP tool	Returns a skill choice to the agent (`skill_select`)	no-decision on no key or HTTP error; 60 s timeout	Tool only. Licence MIT per plugin.json; no LICENSE file	6 Oct · coding-agents row
burnigtm/jev-mcp (`9448f61`); 0.1.0	MCP server (Codex, Cursor)	Returns auto, review or escalate to the agent (`gate` tool)	no-decision on no key or HTTP error; 30 s timeout	Not enforced: the CLI exits non-zero only on errors. Has a mock mode	6 Oct · coding-agents row
minhgv/jev-mcp (`1a1f0a5`); 0.3.0	MCP server and CI shadow CLI (Cursor, Codex)	Returns Jev's answer to the agent; the CI command prints `would_block`	no-decision on no key or HTTP error; SDK default timeout	The CI command always exits 0. Has a mock mode	6 Oct · coding-agents row
pedroknigge/mcp_jev (`32377f8`); 0.0.11	MCP server (Cursor, Codex, Claude and others)	Returns risk signals to the agent	no-decision on no key or HTTP error; 20 s timeout	Its own text: "returns signals only"; the host's allowlist and sandbox are still needed (Reported)	6 Oct · coding-agents row
wellkilo/codex-jev-preflight (`633ddef`)	Codex prompt hook	Adds a task-complexity note to the prompt; gates no tool	advisory on HTTP error it adds an "unavailable" note and the prompt goes on; no key: `not recorded`; 15 s timeout	Read in part	6 Oct · coding-agents row
NiazMorshed2007/jev-review (`57690af`)	MCP server (Codex, Cursor, OpenCode, Claude Code)	Returns code-review quality scores to the agent	no-decision on no key or HTTP error; 30 s timeout, up to 2 retries	Sends the task, diff, full file contents and repository context with no cap or redaction. Not the B row jev-review above (devagrawal09), which is a different project	6 Oct · coding-agents row

Added 18 Oct 2026: 14 grade-A rows

Added 18 Oct 2026: 14 rows, each with the Jev call read at a pinned commit: three Claude Code guards, three game and robot demos, and eight projects (mail and chat filters, camera and photo inspection, a video tagger, a drone-camera demo, a moderation example and a macOS file sorter). Documented The ledger goes from 99 rows (A 86) to 113 (A 100); 86 + 14 = 100. The guards were read on 16 Oct 2026 and their bypass entry points on 18 Oct; the demos on 11 and 12 Oct; the eight projects on 18 Oct 2026, 05:20–05:23 UTC. The pool ran nothing: no hook, app or demo was installed or run, and no Jev call was made. Grade A means the call path was read; none of the 14 is verified as running in production.

What the 14 rows added on 18 Oct do on a Jev error (default settings)

Claude Code guards3 of 14 rows
- fail-openclaude-model-guard: the prompt goes ahead
- heldclaude-jev-advisor: the guard's own ask before a deletion stands
- fail-openohmyjev (lockdown mode): the tool call passes
Game and robot demos3 of 14 rows
- fallbackdoom-jev: last action
- fail-closedjev-robotics-demo: the demo stops
- fallbackJevPlaysSnakeGame: last direction
Projects8 of 14 rows
- 8 of 8error lines described as read; no shared failure term assigned

Guards: fail-open 2, held 1. Demos: fallback 2, fail-closed 1. 3 + 3 + 8 = 14. The guard and demo terms are the ones used on Claude Code and MCP and real-time demos. For the 8 projects the error lines in the code are quoted in the table below and not mapped to a failure term; for jev-content-mod the request is made inside the unread jevlang package. Documented

Claude Code guards · 3 rowsfail-open 2 · held 1 · read 16 Oct; bypass entry points 18 Oct

Full comparison: Claude Code and MCP: which server, hook or plugin fails open? claude-model-guard, claude-jev-advisor and ohmyjev are counted on When Jev fails (80 rows; ohmyjev from 19 Oct 2026, after its pi and omp adapters were read in full).

Project (pinned commit)	Form	What Jev decides	Jev call read at the pin (file and lines)	On a Jev failure	Checked · sources
mcowdery/claude-model-guard (`bced870`); package 1.2.0, not on a registry	Claude Code `UserPromptSubmit` hook	Blocks a prompt when Jev confidently says a bigger model tier is needed (0.65 or more; 0.9 or more within 3 prompts of a model switch). A model-tier gate, not a safety gate. Model `jev-latest`	`scripts/jev.mjs` `callJev` L122–140; `scripts/promptAdvisor.mjs` block L133–144	fail-open on no key, timeout (2,500 ms), HTTP error or malformed answer: the hook exits with no output and the prompt goes ahead. Below its threshold: advisory (a note to Claude)	Read 16 Oct 2026; bypass entry points read 18 Oct 2026 · R17-S242, R17-S243 · counted on the failure map
delt96/claude-jev-advisor (`f2c1791`) = npm `@delt/claude-jev-advisor` 0.5.0; `rm` hook	Claude Code `PreToolUse` hook for Bash and PowerShell (Windows only)	Its own guard asks before a deletion; Jev can only lift that ask, when every target scores 0.8 or more. Model `jev-latest`	`src/jev.ts` `callJev` L41–84; `src/rm/run.ts` `runRmHook` L73–124	held on no key, timeout (8,000 ms), HTTP error or malformed answer: the guard's ask stands. Below 0.8: held	Read 16 Oct 2026; bypass entry points read 18 Oct 2026 · R17-S239, R17-S240 · counted on the failure map
Novacon/ohmyjev (`d307c39`, npm 0.13.2)	Claude Code module: Bash, Write/Edit and exfiltration gates, enforcing deny in lockdown mode (its default is off)	Allows or denies a tool call from the command, the content to write or the fetch and MCP arguments. Model `jev-1.13.0` on TypeSafe (a pinned version)	`hooks/jev.ts` L13–16, L58–64; `hooks/ohmyjev.ts` `askJev` L68–93, gate L270–312 (at `d307c39`)	In lockdown mode: fail-open on no key, timeout (1,500 ms), HTTP error or malformed answer: the call passes. Below its threshold: acts-anyway	Read at `d307c39` 2 Nov 2026 (R26-S20 to R26-S23); earlier at `78bb82c`: read 16 Oct 2026; every entry-point file read for its Jev paths 18–19 Oct 2026 · R17-S236, R17-S237, R19-S200 to R19-S206 · counted on the failure map from 19 Oct 2026 (bypass recorded); see Claude Code and MCP Change, 21 Oct 2026: build/coding-agents now counts 14 of 27 integrations read in source. ohmyjev's pi and omp adapters (this row) joined its table; auto-mode and jcode were read but are not settled and are not ledger rows. Change, 30 Oct 2026: ohmyjev 0.13.2 (29 Oct 2026) removes the auto-approve thresholds from its policy file (R24-S38). The row's grade and failure terms are unchanged. build/coding-agents now counts 14 of 28 integrations read in source as acting on Jev's answer; acmelabs-15/jev-gate was read and is not settled, so it is not a ledger row. Change, 2 Nov 2026: re-read at 0.13.2 (`d307c39`); its gates run only in lockdown mode, which is off by default. Grade and failure terms unchanged; bypass cell changed in content (default lockdown-off skips every gate; auto-approve path removed; GitHub reminder added) (R26-S20 to R26-S23). Noted 7 Nov 2026: ohmyjev 0.14.0 (3 Nov 2026) is newer than the 0.13.2 read; not re-read. Grade and failure terms as read at 0.13.2 (R27-S25).

Game and robot demos · 3 rowsfallback 2 · fail-closed 1 · read 11–12 Oct

Full comparison: Jev plays Doom: what real-time demos do when a Jev call fails. These are demos, not products in use.

Project (pinned commit)	Form	What Jev decides	Jev call read at the pin (file and lines)	On a Jev failure	Checked · sources
mikespins/doom-jev (`bd4e98f`)	Runnable demo (ViZDoom/Freedoom game)	One of 6 actions each tick (0.12 s) through `AsyncTypeSafeClient`; timeout 2.0 s, no retries	`brain.py` L291, L331–334 (lines 1–55 and 280–360 read)	fallback-last action on a Jev error: the game keeps applying the last action. No key: fail-closed (exits before the game starts)	11 Oct 2026 · R13-S68, R13-S102 · real-time demos row
FazalAAli/jev-robotics-demo (`531de61`)	Runnable demo (MuJoCo arm simulation)	The next target, done, and grasp or release at waypoints	`jev_agent.py` L213 (lines 1–40 and 190–308 read)	fail-closed on a Jev error: the demo stops and no action is taken (no `try` in the file)	11 Oct 2026 · R13-S69, R13-S108 · real-time demos row
bileshg/JevPlaysSnakeGame (`b9237bf`)	Runnable demo (Snake game)	The next direction (up, down, left, right) through `TypeSafeClient()`; typesafe-sdk 0.7.4 locked	`src/agents/jev_agent.py` L54–64; `src/game/engine.py` L125–128; `main.py` L12	fallback-last action on a Jev error or no key: the agent returns nothing and the snake keeps its direction	12 Oct 2026 · R14-S49 · real-time demos row

Projects · 8 rowserror lines as read · read 18 Oct

Sources: the project leads recorded on 14 and 16 Oct 2026 and one Show HN post (Putaway). Two other reads did not become rows: Koh-Henry/JevMail (no Jev call in the repository at 2f14d18) and romantic28/jevGameAgent (no file mentions Jev at ad3961e).

Project (pinned commit)	Form	What Jev decides	Jev call read at the pin (file and lines)	On a Jev failure	Checked · sources
soderlind/jev-mail-triage (`29cebd0`)	CLI and service for Proton Mail Bridge	Spam and scam Noul scores; at the threshold the message moves to review, otherwise it stays. Moving is off by default (`config.py` L191)	`src/jev_mail_triage/classifier.py` L31; `policy.py` L30–33	A Jev error or invalid answer raises `RetryableError` (`classifier.py` L73–82). When a local gate is enforced, its verdict replaces the Jev call (`service.py` L228–232)	18 Oct 2026, 05:21 UTC · R18-S120 to R18-S123
igrady2/school-chat-filter (`c664349`)	School chat moderation web app	Per-category Noul scores decide what happens to a message	`src/lib/moderation/typesafe.ts` L85–92; `index.ts` L105–111	On a provider error the message is not posted (`post.ts` L189–205: 502 on a rejected key, otherwise 503). Timeout 15,000 ms (`typesafe.ts` L92)	18 Oct 2026, 05:21 UTC · R18-S124 to R18-S126
seominpapa/sitesee (`d7fb822`)	Construction-site camera safety pipeline (YOLO plus Jev)	Hazard probability, risk score and action plan from a YOLO text summary; images are sent only on fallback	`siteeye/jev_client.py` L138–160; `pipeline.py` L95	A transport failure after retries raises `RuntimeError` (`jev_client.py` L149–160). How the caller handles it: not recorded	18 Oct 2026, 05:21 UTC · R18-S127, R18-S128
MrCull/QuaySight (`c9fd664`)	Shipping-container photo inspection web app	Container details and damage read from photos (image input; Jev 1.14 for Box). The user supplies their own key in the browser	`src/services/jev.ts` L59–66; `src/domain/jev.ts` L25	A network failure raises `JevError` (`jev.ts` L67–71). Where the readings go next (export) was not traced	18 Oct 2026, 05:21 UTC · R18-S129, R18-S130
botheaj/video-image-tagger (`e628367`)	Rust CLI that finds a product in video frames	A "product visible" Noul at or above the threshold counts as a match; outlined and cropped images are saved (`main.rs` L151)	`src/jev.rs` L174–178 (POST), L94–98; `src/main.rs` L139	Any frame error ends the run with exit code 2 (`main.rs` L81–88, L121–131)	18 Oct 2026, 05:21 UTC · R18-S131, R18-S132
twidtwid/jev-dronecam (`236227f`)	Browser demo with a simulated drone camera	Box plus a "seen" Noul gate; above the threshold the simulated drone approaches	`server.py` L78–83; `public/index.html` L381–382	HTTP error → 502, unreachable → 504 (`server.py` L125–128); the page retries and stops after 4 failures (`index.html` L379). The demo URL answered HTTP 200 at 05:22 UTC (page loaded only)	18 Oct 2026, 05:21 UTC · R18-S133 to R18-S135
kallurayaankit/jev-content-mod (`90f644b`)	Content moderation example with asymmetric thresholds	Harm, category, clarity and needs-human scores → approve, shadow, remove or send to a person (`policy.py` L50–94; needs-human escalates, L85–89)	`moderate.jev` L1–20 (`ask_all`); `policy.py` L40–46	The HTTP request to Jev is made inside the third-party `jevlang` package, which was not read, so its error path is not recorded. Graded A on the call site in the repository, as for rows whose call goes through an SDK	18 Oct 2026, 05:23 UTC · R18-S136, R18-S137
devYRPauli/putaway (`35d8c50`)	macOS app that files a folder's contents into existing folders	The folder for each file; a text file at or above the auto-move confidence is marked ready to move, otherwise the user is asked. Per the README nothing moves until the user clicks Move	`Sources/PutawayCore/Classify.swift` L62, L81–87; `Plan.swift` L59–62	401/403 and 402 → `ClassifierUnavailable` (`Classify.swift` L113–131). SubQ is an alternative provider (L61)	18 Oct 2026, 05:23 UTC · R18-S138, R18-S139

Most rows were discovered between 20 and 25 September; QuantDinger was discovered on 27 September. The discovery date is not the date a project was created. The 24 rows added on 3 Oct were read at pinned commits between 1 and 3 Oct. The 6 rows added on 4 Oct, and jev-trader and Jev Trade (listed since 25 and 27 Sep), were read at pinned commits on 4 Oct. The 5 rows added on 5 Oct, the 3 moved from B (listed since 25 Sep) and jev-ultrafast were read at pinned commits on 5 Oct. The 20 rows added on 6 Oct were read at pinned commits on 6 Oct (the 2 guard rows at their 29 Sep commits, unchanged). The 2 rows added on 10 Oct were read at a pinned commit and a pinned npm version on 10 Oct. The 3 rows added on 11 Oct were read at pinned commits and a pinned npm version on 11 Oct. Of the 14 rows added on 18 Oct, the 3 guards were read at pinned commits on 16 Oct (bypass entry points on 18 Oct), the 3 demos on 11 and 12 Oct and the 8 projects on 18 Oct. Links to main branches can change after the check date.

Production claims by companies

Companies and analysts have published their own statements about using Jev. None has been independently verified; each is the claimant's own statement, checked against its source on 28 Sep 2026 and rechecked on 3 Oct 2026 and 18 Oct 2026, 05:19 UTC. Reported

5 of 7 rechecked 18 Oct: online, same text2 of 7 not rechecked: relayed by VentureBeat (HTTP 429)0 of 7 changed

Rechecked 18 Oct 2026, 05:19 UTC: Brizz, Polylane, the Vercel CEO post, the Bryo AI post and FundaAI are online with the same text; the two figures relayed by VentureBeat were not rechecked (no direct link; VentureBeat returned HTTP 429). Reported

Claimant	Claim as worded	What the source actually shows	Source and date	Independent check	Recheck, 18 Oct 2026 (UTC)
Brizz	"21x cheaper"; step spend "drops from $11,594 to $2,759, or 76%"	A measured backfill of 17,840 requests: $1.40 on Jev vs $29.67 on Gemini; mean latency 3.0 s vs 49.0 s. The 76% is a projection on one customer's traffic, not an observed invoice.	Company blog, 23 Sep	None	Online (HTTP 200); same text; no newer Jev post in the blog index · 05:19:19 (3 Oct: unchanged)
Polylane	"We swapped our LLMs for Jev. It's 39% cheaper"; "we put it in production"	About one week of comparison. Estimated cost per 1,000 calls fell from $0.76199 to $0.46369; P90 latency from 4,752 ms to 508 ms.	Company blog, 25 Sep; Hacker News post 28 Sep, 17:37 UTC	None	Online (HTTP 200); same text; no newer Jev post in the blog index · 05:19:19 (3 Oct: unchanged)
Vercel (command safety in `fx`)	Reported elsewhere as a completed swap	Benchmarked Jev for command safety. On 16 Sep the CEO wrote that the reviewer "runs on GPT Luna today", that Jev "is up to 18x faster (p95) and more accurate", and that it is "likely new default". A benchmark and an intent, not a completed swap. Test set and size are not public. Benchmark part graded D on the benchmarks ledger (B26).	CEO post on X, 16 Sep	None	Online (X oEmbed 200); same text; a newer statement could not be checked (no X search) · 05:19:19 (3 Oct: same text)
Vercel (adoption)	"roughly 13% of its paid AI Gateway teams were running it within 24 hours"	A company statement relayed by VentureBeat. Vercel's own post "Open weight takes 71.5% of tokens and 23% of spend" (AI Gateway production index, October 2026; read 19 Oct 2026, 05:35 UTC) states a different figure: Jev reached "35% of paid teams in its first two weeks" (35.3% in its reach table) but "accounted for just 0.27% of gateway spend over that same period"; of the teams that tried Jev in week one, 55.6% returned in week two. It covers paid AI Gateway teams only; Vercel sells gateway access to Jev; the post's publication date was not read on the page. Reported R19-S64	VentureBeat, 21 Sep (read through a summarising tool; direct access returned HTTP 429); Vercel blog (production index)	None	Not rechecked 18 Oct 2026 (VentureBeat returned HTTP 429): this page has no direct link to the VentureBeat article (not rechecked on 3 Oct either)
TypeSafe	"cleared 140,000 people from the waitlist within 36 hours"	A vendor figure relayed by VentureBeat.	VentureBeat, 21 Sep (same access limit)	None	Not rechecked 18 Oct 2026 (VentureBeat returned HTTP 429): this page has no direct link to the VentureBeat article (not rechecked on 3 Oct either)
Bryo AI (CTO)	"Jev lost to Gemini on our email classification benchmark. I'm still interested in putting it into production."	An offline benchmark on 1,565 German and English business emails in 10 categories. Not a production claim. Graded C on the benchmarks ledger (B12); the figures 96.4% vs 97.5–98.5% come from Prefactor, a secondary source.	Post on X, 17 Sep	None	Online (X oEmbed 200); same text; a newer statement could not be checked (no X search) · 05:19:19 (3 Oct: same text)
FundaAI	"80 of 2,153 hands-on cases (3.7%) are in production"	Their own classification of public launch-week posts; paid article, method and data not inspected.	Substack analysis, 24 Sep	None	Online (HTTP 200); same text (published and modified 24 Sep 2026) · 05:19:19 (3 Oct: unchanged)

Left out: a claim attributed to Notra, because no primary source was found and the VentureBeat articles checked do not mention it. TechCrunch's Vercel article was not checked. No new production claim appeared in the titles of 15 Jev stories on Hacker News read in the 24-hour scan before 3 Oct 2026, 05:23 UTC (titles only). Bounded finding, re-dated 18 Oct 2026: no production deployment in the 113 ledger rows or the 7 claims is independently verified.

Are the earlier grade-A projects still online?

Checked 3 Oct 2026, 05:23 UTC, for the 15 rows graded A through 29 Sep. Documented

14 GitHub repositories, all public, none archived: QuantDinger, Jevmail, Jev Comment Triage, sqlite3-jev, jev-ultrafast, Canny, killmyidea, json-render, typesafe-router, jkudish/jev-mcp, danna-zhou/jev-mcp, jev-claude, oh-my-claudecode and typesafe-ai/skills each answered HTTP 200 with no "archived" banner (05:23:00–05:23:01 UTC). This was a check of the repositories' HTML pages, not the GitHub API (used to save API quota).
The 15th row, LangChain langchain-typesafe, is a PyPI package: still 0.0.1a3, unchanged.
No grade-A repository was deleted or archived. The 24 rows added on 3 Oct were read at their pinned commits between 1 and 3 Oct and were not checked again for this list.

Listed elsewhere but not counted as products

JevCases, Made with Jev, awesome-jev-cobanov, awesome-jev-use-cases, typesafeai.app and OpenRouter's model-page Apps list are discovery lists. Jev in the Wild reports a 2,170-project dataset; that is a count of projects found, not 2,170 verified products.

Counts shown by the lists on 18 Oct 2026, 05:19 UTC: Made with Jev 855 builds ("Updated Oct 18, 2026"); typesafeai.app 163 records (its page says "Zero records have been independently reproduced"); JevCases 44 indexed cases; Jev in the Wild 2,170 (unchanged, v1 only). awesome-jev-cobanov, awesome-jev-use-cases and OpenRouter's Apps list show no total. Directory counts are listings, not adoption. Two newer awesome lists seen on 16 Oct (reacher-z/Awesome-Jev, AwesomeJev/AwesomeJev.github.io) were not opened. Reported

LiteLLM's complexity router with classifier_type: "jev" (BerriAI/litellm at 8d28e8d, read 2 Oct 2026) is a framework feature, not a product, so it is not counted. On a Jev error it falls back to a local scorer by default. It is compared on model routing. The LiteLLM compaction guardrail (version 1.104.0), counted on the failure map since 8 Oct, is also a framework feature and is not counted here. Documented

Two judge libraries are handled the same way: DeepEval JevEval (version 4.2.8, commit a200ece; deepeval 4.3.3 (4 Nov 2026) is newer than the version this page describes; not re-read, R27-S25), an eval metric, and openlayer jevals (commit 0a8f895), an eval and runtime-gate library. Both are library features, not products, so they are not counted. Read 4 Oct 2026: on a Jev error DeepEval stops the run by default; openlayer jevals leaves the item out of the score, and its report still counts it as passed. They are compared on Jev as a judge. Documented

Seven Jev framework packages are also framework features, not products, so they are not counted: langchain-typesafe 0.0.1a3 (LangChain), @langchain/typesafe 0.0.2 (LangChain), the Pydantic AI TypeSafeModel (in pydantic-ai-slim 2.54.0), @effect/ai-typesafe 4.0.1 (Effect), @tanstack/ai-typesafe 1.0.0 (TanStack), the community typesafe-spring-ai 0.4.0 (org.springaicommunity, not part of spring-projects/spring-ai) and the community llama-index-postprocessor-jev 0.1.1. Read at pinned tags on 5 Oct 2026 (TanStack AI at commit 5651e07 on 31 Oct 2026): 4 of the 7 are pre-1.0 and none is published by TypeSafe; 5 raise Jev errors to the caller; the Spring AI passage filter fails open and the LlamaIndex reranker falls back to the retriever's order (the package calls this "fail open"). Change, 31 Oct 2026: TanStack AI added (it raises to the caller, so it is not a product row); this paragraph said six packages, 4 of the 6 pre-1.0 and 4 raising errors, and that the LlamaIndex reranker fails open. The LangChain AutoModeMiddleware row in the integrations table above stays counted, as it was before 5 Oct. Its JS twin, @langchain/typesafe 0.0.2, has the same shape but is not counted: the Python row was counted before the framework rule of 5 Oct and both are kept as published (editorial decision, 18 Oct 2026). They are compared on Jev LangChain and other framework integrations. Documented

What was not verified

No project was observed making a live Jev call, and no installation, user count or bill was checked. In this refresh nothing was installed or run and no mailbox or workflow was connected.
Grade A means the call path was read at a commit, not that the project runs, has users or works well.
Grade B rows were not checked against their code in the run that recorded them. None was re-read in source between 1 and 3 Oct; on 4 Oct two (jev-trader and Jev Trade) were read and moved to A; on 5 Oct three (Jev for Chrome, computer-use-jev and jev-android) were read and moved to A. The other 16 were not re-read.
For the 2 rows of 10 Oct (jev-effort-router, jev-gateway), no agent, plugin or proxy was installed or run and no Jev call was made; whether either is used in production: not verified.
For the 20 rows of 6 Oct, no agent, hook, plugin or MCP server was installed or run and no Jev call was made. Not read: what Codex, Cursor, OpenClaw and OpenCode do after a hook or plugin answers (Hermes core's approval path and OpenCode's default rules were read on 7 Oct); rh-guard's redaction; whether the osuki 0.2.7 and 0.2.8 tarballs match their tags file by file; test file contents. Whether any of them is used in production: not verified.
For the 8 rows of 5 Oct and jev-ultrafast, no agent was run, no extension was installed and no key was used. Test contents were read for few of them. Not recorded: whether computer-use-jev sends [secure] field values, second-look's fallback card wording, and trycua/cua's no-key behaviour. jev-agent-browser's npm tarball was not compared to its repository. The AndroidWorld result in arXiv 2609.30186 (single author, submitted 24 Sep 2026) is the author's claim (Reported); the pool states no figure.
For the 8 rows of 4 Oct and QuantDinger, no bot, order, payment, eval or Jev call was run. Jev Trade's live-fill claims are the creator's own. Not traced: jev-trader's no-key path (inside @ai-sdk/typesafe-ai) and whether QuantDinger's worker upgrade of queued orders from virtual to live passes the filter. competitor-hunter loads whatever Jev caller the user has installed, so that version is not pinned. did-they-answer's stop on a malformed answer was traced in source, not run.
For the 14 rows of 18 Oct, no hook, app or demo was installed or run and no Jev call was made. Not read: the jevlang package that makes jev-content-mod's request; ohmyjev's pi and omp adapters beyond their tool-call gates (read in full on 19 Oct 2026); how sitesee's caller handles an error; where QuaySight's readings go next. For jev-dronecam the demo page was loaded only. Whether any of the 14 is used in production: not verified.
Creator timing, cost and accuracy figures in project READMEs are their own reports.
Rows keep their own check dates (23 Sep to 18 Oct). Repository status on 3 Oct was read from GitHub HTML pages, not the API.
Production claims were checked against their own sources only; no company confirmed or invoiced anything to the pool. The two claims relayed by VentureBeat (Vercel's 13% adoption figure and TypeSafe's 140,000 cleared from the waitlist) were not rechecked on 3 Oct or on 18 Oct 2026 (VentureBeat returned HTTP 429).

Method, sources and check times (UTC)

Changed on 2 Nov 2026: the ohmyjev row (Claude Code guards) re-read at d307c39 (0.13.2) and names lockdown mode (default off). ohmyjev (d307c399fbc4898200fb2c229c623df9a517f318, 0.13.2; R26-S20 to R26-S23, read 2 Nov 2026; earlier 78bb82c, R17-S236, R17-S237, 16 Oct 2026). No row, grade or count changed.
Changed on 30 Oct 2026: the ohmyjev row (Claude Code guards) notes that ohmyjev 0.13.2 (29 Oct 2026) removes the auto-approve thresholds from its policy file; the row was read at an earlier version and has not been re-read at 0.13.2 (policy files of 0.12.0 and 0.13.2 compared, 30 Oct 2026, 05:20 UTC; R24-S38). No row, grade or count of this ledger changed.
Changed on 24 Oct 2026: wording only: the method lines for the rows of 4 and 5 Oct now say the key lines were spot-checked against the raw files at the same commits. No row, grade or count changed.
Changed on 18 Oct 2026: 14 new A rows. Claude Code guards: claude-model-guard (bced870dd65fae00865964b044f1a9717f917a85; R17-S242, R17-S243), claude-jev-advisor (f2c17914a591c443e58c7a4b260a40d25e261fd6; R17-S239, R17-S240) and ohmyjev (78bb82cea03c0fb06fbf8892dc6d1840a7f02c11; R17-S236, R17-S237), read 16 Oct 2026, 05:23–05:25 UTC; HEAD equal to the pin on 18 Oct, 05:20 UTC. Demos: doom-jev (R13-S68, R13-S102), jev-robotics-demo (R13-S69, R13-S108), JevPlaysSnakeGame (R14-S49), read 11 and 12 Oct. Projects: 8 repositories read at full SHAs on 18 Oct 2026, 05:20–05:23 UTC (R18-S120 to R18-S139); each pin came from the repository's commits feed. 113 rows (A 100, B 12, C 1): A 86 + 14 = 100; 100 + 12 + 1 = 113. No row changed grade. Independently verified in production: still 0 of 113. Recount from this page's rows and cross-check against the failure map and the Claude Code page: R18-S100 to R18-S102. Claims rechecked 05:19 UTC (R18-S103 to R18-S111); discovery-list counts 05:19 UTC (R18-S112 to R18-S119). The coding-agent figure's outcome grouping was replaced to match the failure terms (counts unchanged).
Changed on 11 Oct 2026: 3 new A rows: jev-triage (6360e21; host OpenClaw aa6008ad198ef99c43f9d89dbd01694708712974; R13-S200 to R13-S207), the jev-for-all launcher (npm 0.7.1, gitHead 74f62e71ee12ae29ce278c6e1005feec0e78de46; R13-S222 to R13-S224) and hermes-jev-mana (3daf639; R13-S210 to R13-S216); all read 11 Oct 2026, 05:17–05:27 UTC. 99 rows (A 86, B 12, C 1): A 83 + 3 = 86; 86 + 12 + 1 = 99. No row changed grade. jev-approvals and jev-effort-router wording updated (R13-S217). Per-row failure cells and line ranges on build/coding-agents. Independently verified in production: still 0 of 99.
Changed on 10 Oct 2026: 2 new A rows: jev-effort-router (Hermes plugin, 562dc2a, v0.2.2; R12-S200 to R12-S209) and jev-gateway (npm 0.5.1, gitHead 37ff982949f0cb9bb131963f82a6f34f8ba38edd; R12-S257); both read 10 Oct 2026. 96 rows (A 83, B 12, C 1): A 81 + 2 = 83; 83 + 12 + 1 = 96. No row changed grade. Per-row failure cells on build/coding-agents. Independently verified in production: still 0 of 96.
Changed on 8 Oct 2026: 2 new A rows: LanceDB TypeSafeReranker in v0.40.0 (tag object 2fd3239 → commit b1b080b; python/python/lancedb/rerankers/typesafe.py L43–46, L67–69, L125–127, L158–177, L186, L199–203; R11-S204, R11-S205, read 05:13:40 UTC) and pg-redact at ee903ef (src/lib/jev.ts L5, L34–46, L55, L64–75; src/app/api/classify/route.ts L29–58; R11-S220, R11-S221, read 05:16:00 UTC). jev-reranker (Python) moved from B to A: PyPI 0.1.2 (PyPI JSON, R11-S203), tag v0.1.2 = d58594b; src/jev_reranker/reranker.py L126–127, L233–235, L512–517, L546–552; src/jev_reranker/_client.py L44–67, L100–114, L132–136 (R11-S200 to R11-S202, read 05:13:20 UTC). The ledger went from 92 rows (A 75, B 16, C 1) to 94 (A 81, B 12, C 1): A 75 + 2 new + 4 from B = 81; B 16 − 4 = 12; 81 + 12 + 1 = 94. pg-jev, sqlite3-jev, sqlite-jev and DocJev rows now link their failure cells on Jev inside SQL and document pipelines; pg-jev, sqlite-jev and DocJev moved from B to A because their call paths were read at pinned commits (sqlite3-jev was already A). Version lines: oh-my-claudecode checked v5.6.2 (454bae0; R11-S100 to R11-S102) and jkudish/jev-mcp checked 0.14.1 (R11-S103, R11-S104), failure paths unchanged. hermes-jev-approvals Jev-error cell corrected to fallback-main agent model (held on other HTTP errors) (R11-S110 to R11-S112); hermes-jev-curator no-key, malformed and below-threshold cells filled (R11-S108, R11-S109). The coding-agent row counts (8 of 18 act on Jev's answer, 10 of 18 not counted) are unchanged; on a Jev error 5 of the 8 can now let the tool call run (was 4) and 3 of 8 hold it (was 4). Sources: the 8 Oct research report. Independently verified in production: still 0 of 94.
Changed on 6 Oct 2026: 20 new A rows: 18 coding-agent integrations (keiffff/jev-kit, bloudhood/codex-jev-approval-hook, 24601/rh-guard, justmytwospence/opencode-tool-gate, @osuki-dev/opencode-osuki-agent 0.2.7, DoGMaTiiC/hermes-jev, anpicasso/hermes-jev-approvals, anpicasso/hermes-jev-curator, DECRUX9812/typesafe-skill-router, ourines/hermes-jev, ajensenwaud/hermes-jev-plugin, ydmw74/jev-skill-router, @openclaw/typesafe 2026.9.8, burnigtm/jev-mcp, minhgv/jev-mcp, pedroknigge/mcp_jev, wellkilo/codex-jev-preflight, NiazMorshed2007/jev-review) and 2 guard rows (madisonrickert/jev-permission-gate, gulbaki/jev-llm-guard). The ledger went from 72 rows (A 55, B 16, C 1) to 92 (A 75, B 16, C 1); recounted 6 Oct 2026 from the rows on this page. Independently verified in production: still 0 of 92. The existing B rows "jev-review" (devagrawal09) and "openclaw-typesafe-ai" (Olli0103) are different projects and stay. Claims and repository status were not rechecked on 6 Oct.
Rows of 6 Oct: source read at the pinned commit linked in each row, 6 Oct 2026, 05:14–05:29 UTC; HEAD of the two guard repositories checked at 05:20 UTC. The two npm packages were downloaded, read and deleted; nothing was run. Per-row paths, line numbers and source IDs are on build/coding-agents and build/claude-code-mcp.
Changed on 5 Oct 2026: 5 new A rows in browser and computer use (ironbee-express, jev-agent-browser, Ulka, second-look, trycua/cua jev-use, which is mock by default); Jev for Chrome, computer-use-jev and jev-android moved from B to A; jev-ultrafast re-read at 1231850; six framework packages listed as not counted. The ledger went from 67 rows (A 47, B 19, C 1) to 72 (A 55, B 16, C 1). Independently verified in production: still 0. Claims and repository status were not rechecked on 5 Oct.
Rows of 5 Oct: source read at the pinned commit linked in each row, 5 Oct 2026, 05:12–05:24 UTC; 15 key lines were spot-checked against the raw files at the same commits (14 matched as quoted; one line reference corrected). Details on browser and computer use. Framework packages: pinned tags read the same morning; details on framework integrations. The pool's 5 Oct research report lists the corrections of that run.
Changed on 4 Oct 2026: 6 new A rows (btc-hft-jev, jev-fraud-shield, fraud-jev, the expense-policy gate, did-they-answer, competitor-hunter); jev-trader and Jev Trade moved from B to A; QuantDinger re-read at a new commit, with its failure behaviour and order call sites updated; DeepEval JevEval and openlayer jevals listed as not counted. The ledger went from 61 rows (A 39, B 21, C 1) to 67 (A 47, B 19, C 1). Independently verified in production: still 0. Claims and repository status were not rechecked on 4 Oct.
Rows of 4 Oct: source read at the pinned commit linked in each row, 4 Oct 2026, 05:14–05:21 UTC, with spot checks against the raw files at the same commits (05:20–05:26); details on trading and fraud gates and Jev as a judge. competitor-hunter's Jev caller was read at claude-x-jev c8662b0, but the copy a user has installed is not pinned.
QuantDinger: re-read on 4 Oct, 05:17–05:20 UTC, at a5a9f4c (1 Oct). backend_api_python/app/services/ai_decision_filter.py was rewritten after the 28 Sep audit (now 743 lines), so line links from that audit no longer apply. Pinned lines: Jev error, then LLM, then allow (L194–238); no key (L156–175); confidence below 0.55 (L285–293); malformed answer (L437–468); allow rule (L305–309). Order call sites: strategy_v2/live_execution.py L233–273, routes/quick_trade.py L564–567 and routes/alpaca.py L453–479.
New rows (24): source read at the pinned commit linked in each row, on the analysis pages: support-ticket triage (1 Oct 2026), model routing and content moderation (2 Oct), email and lead sorting and n8n nodes (3 Oct). Duplicate check against this ledger on 3 Oct, 05:23: none of the 24 was already listed.
Jevmail: re-read on 3 Oct at f6f20af (last commit 19 Sep; classifier unchanged since the 28 Sep audit). Jev Comment Triage: pinned at d848a5b on 2 Oct.
oh-my-claudecode: release v5.6.0 (published 1 Oct 2026, 05:18 UTC), commit e74b22c; src/hooks/jev/config.ts L36, resolver.ts L10–12, read 3 Oct, 05:20–05:22: rechecked at v5.6.0 (e74b22c), failure path unchanged. The row's verdict comes from the v5.5.0 audit at 9fd35ec.
Claims rechecked 3 Oct: Brizz (05:23:01), Polylane (05:23:02), Vercel CEO post via X oEmbed (05:23:03), Bryo AI CTO post via X oEmbed (05:23:03), FundaAI (05:23:03). The two VentureBeat-relayed claims were not rechecked.
Repository status: GitHub repository HTML pages for the 14 grade-A repositories (05:23:00–05:23:01); langchain-typesafe on PyPI.
Page title: autocomplete from Google (client=firefox), YouTube, Bing and DuckDuckGo, 3 Oct, 05:22:12–05:22:14. "jev examples" was offered by all four; "built with jev" by Google; typing "jev products" brought "jev projects" on Bing and DuckDuckGo, but "jev products" itself was not suggested; "apps using jev" gave no Jev suggestion. These are demand signals, not search volumes.
Earlier rows: audits of 23, 25, 28 and 29 Sep 2026 (reports linked below).